# PoP - Policing Socio-Geographic Boundaries and Inequality
# script for creating Figure 5
# Figure 5 - Heterogeneous Influence of Racial Boundary on Arrests By White Racial Context. 


# load packages
suppressPackageStartupMessages(
  
  {
    library(dplyr)
    library(tidyverse)
    library(ggplot2)
    library(haven)
    library(readxl)
    library(readr)
    library(areal)
    library(car)
    library(estimatr)
    library(magrittr)
    library(texreg)
    library(sandwich)
    library(ggh4x)
    library(jtools)
  }
)

# load and format all data
# First ATL
# Load Atlanta final dataset
load("atl_final.RData")

# Set up DVs
# log and +1 to variables (pop already logged)
atl_fin = atl_fin %>% 
  mutate(lmhhi = log(mhhi + 1),
         larrests = log(total_arrests + 1),
         lmisdemeanors = log(misdemeanor_arrests + 1),
         lfelonies = log(felony_arrests + 1),
         lnonviolent = log(nonviolent_arrests + 1),
         lviolent = log(violent_arrests + 1),
         lsociety = log(society_arrests + 1),
         lperson = log(person_arrests + 1),
         lproperty = log(property_arrests + 1),
         lcrime = log(crime_all + 1),
         lpropertycrime = log(crime_property + 1),
         lviolentcrime = log(crime_violent + 1))

# create scaled DVs (to account for different pop. sizes and density in blocks)
atl_fin$larrest_sd <- atl_fin$larrests - (mean(atl_fin$larrests)/sd(atl_fin$larrests))
atl_fin$lmisdemeanors_sd <- atl_fin$lmisdemeanors - (mean(atl_fin$lmisdemeanors)/sd(atl_fin$lmisdemeanors))
atl_fin$lfelonies_sd <- atl_fin$lfelonies - (mean(atl_fin$lfelonies)/sd(atl_fin$lfelonies))
atl_fin$lnonviolent_sd <- atl_fin$lnonviolent - (mean(atl_fin$lnonviolent)/sd(atl_fin$lnonviolent))
atl_fin$lviolent_sd <- atl_fin$lviolent - (mean(atl_fin$lviolent)/sd(atl_fin$lviolent))
atl_fin$lsociety_sd <- atl_fin$lsociety - (mean(atl_fin$lsociety)/sd(atl_fin$lsociety))
atl_fin$lperson_sd <- atl_fin$lperson - (mean(atl_fin$lperson)/sd(atl_fin$lperson))
atl_fin$lproperty_sd <- atl_fin$lproperty - (mean(atl_fin$lproperty)/sd(atl_fin$lproperty))

# now create binned racial blv measures
atl_fin <- atl_fin %>% mutate(boundary_quart = quantile(atl_fin$p_race_white_blv, prob=.75, na.rm=TRUE),
                              boundary_quart_dummy = ifelse(p_race_white_blv > boundary_quart,1,0 ))


# For Austin
# Load Austin final dataset
load("aus_final.RData")


# Set up DVs
# log and +1 to variables (pop already logged)
aus_fin = aus_fin %>% 
  mutate(lmhhi = log(mhhi + 1),
         larrests = log(total_arrests + 1),
         lmisdemeanors = log(misdemeanor_arrests + 1),
         lfelonies = log(felony_arrests + 1),
         lnonviolent = log(nonviolent_arrests + 1),
         lviolent = log(violent_arrests + 1),
         lsociety = log(society_arrests + 1),
         lperson = log(person_arrests + 1),
         lproperty = log(property_arrests + 1),
         lcrime = log(crime_all + 1),
         lpropertycrime = log(crime_property + 1),
         lviolentcrime = log(crime_violent + 1))

# create scaled DVs (to account for different pop. sizes and density in blocks)
aus_fin$larrest_sd <- aus_fin$larrests - (mean(aus_fin$larrests)/sd(aus_fin$larrests))
aus_fin$lmisdemeanors_sd <- aus_fin$lmisdemeanors - (mean(aus_fin$lmisdemeanors)/sd(aus_fin$lmisdemeanors))
aus_fin$lfelonies_sd <- aus_fin$lfelonies - (mean(aus_fin$lfelonies)/sd(aus_fin$lfelonies))
aus_fin$lnonviolent_sd <- aus_fin$lnonviolent - (mean(aus_fin$lnonviolent)/sd(aus_fin$lnonviolent))
aus_fin$lviolent_sd <- aus_fin$lviolent - (mean(aus_fin$lviolent)/sd(aus_fin$lviolent))
aus_fin$lsociety_sd <- aus_fin$lsociety - (mean(aus_fin$lsociety)/sd(aus_fin$lsociety))
aus_fin$lperson_sd <- aus_fin$lperson - (mean(aus_fin$lperson)/sd(aus_fin$lperson))
aus_fin$lproperty_sd <- aus_fin$lproperty - (mean(aus_fin$lproperty)/sd(aus_fin$lproperty))

# now create binned racial blv measures
aus_fin <- aus_fin %>% mutate(boundary_quart = quantile(aus_fin$p_race_white_blv, prob=.75, na.rm=TRUE),
                              boundary_quart_dummy = ifelse(p_race_white_blv > boundary_quart,1,0 ))


# For Boston
## Load Boston data
load("bos_final.RData")


# Set up DVs
# log and +1 to variables (pop already logged)
bos_fin = bos_fin %>% 
  mutate(lmhhi = log(mhhi + 1),
         larrests = log(total_arrests + 1),
         lmisdemeanors = log(misdemeanor_arrests + 1),
         lfelonies = log(felony_arrests + 1),
         lnonviolent = log(nonviolent_arrests + 1),
         lviolent = log(violent_arrests + 1),
         lsociety = log(society_arrests + 1),
         lperson = log(person_arrests + 1),
         lproperty = log(property_arrests + 1),
         lcrime = log(crime + 1))



# create scaled DVs (to account for different pop. sizes and density in blocks)
bos_fin$larrests_sd <- bos_fin$larrests - (mean(bos_fin$larrests)/sd(bos_fin$larrests))
bos_fin$lmisdemeanors_sd <- bos_fin$lmisdemeanors - (mean(bos_fin$lmisdemeanors)/sd(bos_fin$lmisdemeanors))
bos_fin$lfelonies_sd <- bos_fin$lfelonies - (mean(bos_fin$lfelonies)/sd(bos_fin$lfelonies))
bos_fin$lnonviolent_sd <- bos_fin$lnonviolent - (mean(bos_fin$lnonviolent)/sd(bos_fin$lnonviolent))
bos_fin$lviolent_sd <- bos_fin$lviolent - (mean(bos_fin$lviolent)/sd(bos_fin$lviolent))
bos_fin$lsociety_sd <- bos_fin$lsociety - (mean(bos_fin$lsociety)/sd(bos_fin$lsociety))
bos_fin$lperson_sd <- bos_fin$lperson - (mean(bos_fin$lperson)/sd(bos_fin$lperson))
bos_fin$lproperty_sd <- bos_fin$lproperty - (mean(bos_fin$lproperty)/sd(bos_fin$lproperty))


# now create binned racial blv measures
bos_fin <- bos_fin %>% mutate(boundary_quart = quantile(bos_fin$p_race_white_blv, prob=.75, na.rm=TRUE),
                              boundary_quart_dummy = ifelse(p_race_white_blv > boundary_quart,1,0 ))




# For Chicago
## Load Chicago data
load("chi_final.RData")


# Set up DVs
# log and +1 to variables (pop already logged)
chi_fin = chi_fin %>% 
  mutate(lmhhi = log(mhhi + 1),
         larrests = log(total_arrests + 1),
         lmisdemeanors = log(misdemeanor_arrests + 1),
         lfelonies = log(felony_arrests + 1),
         lnonviolent = log(nonviolent_arrests + 1),
         lviolent = log(violent_arrests + 1),
         lsociety = log(society_arrests + 1),
         lperson = log(person_arrests + 1),
         lproperty = log(property_arrests + 1),
         lcrime = log(crime_all + 1),
         lpropertycrime = log(property_crime + 1),
         lviolentcrime = log(violent_crime + 1))


# create scaled DVs (to account for different pop. sizes and density in blocks)
chi_fin$larrest_sd <- chi_fin$larrests - (mean(chi_fin$larrests)/sd(chi_fin$larrests))
chi_fin$lmisdemeanors_sd <- chi_fin$lmisdemeanors - (mean(chi_fin$lmisdemeanors)/sd(chi_fin$lmisdemeanors))
chi_fin$lfelonies_sd <- chi_fin$lfelonies - (mean(chi_fin$lfelonies)/sd(chi_fin$lfelonies))
chi_fin$lnonviolent_sd <- chi_fin$lnonviolent - (mean(chi_fin$lnonviolent)/sd(chi_fin$lnonviolent))
chi_fin$lviolent_sd <- chi_fin$lviolent - (mean(chi_fin$lviolent)/sd(chi_fin$lviolent))
chi_fin$lsociety_sd <- chi_fin$lsociety - (mean(chi_fin$lsociety)/sd(chi_fin$lsociety))
chi_fin$lperson_sd <- chi_fin$lperson - (mean(chi_fin$lperson)/sd(chi_fin$lperson))
chi_fin$lproperty_sd <- chi_fin$lproperty - (mean(chi_fin$lproperty)/sd(chi_fin$lproperty))

# now create binned racial blv measures
chi_fin <- chi_fin %>% mutate(boundary_quart = quantile(chi_fin$p_race_white_blv, prob=.75, na.rm=TRUE),
                              boundary_quart_dummy = ifelse(p_race_white_blv > boundary_quart,1,0 ))




# For Milwaukee
## Load Milwaukee data
load("mil_final.RData")


# Set up DVs
# log and +1 to variables (pop already logged) 
mil_fin = mil_fin %>% 
  mutate(lmhhi = log(mhhi + 1),
         larrests = log(total_arrests + 1),
         lmisdemeanors = log(misdemeanor_arrests + 1),
         lfelonies = log(felony_arrests + 1),
         lnonviolent = log(nonviolent_arrests + 1),
         lviolent = log(violent_arrests + 1),
         lsociety = log(society_arrests + 1),
         lperson = log(person_arrests + 1),
         lproperty = log(property_arrests + 1),
         lcrime = log(crime_all + 1),
         lpropertycrime = log(crime_property + 1),
         lviolentcrime = log(crime_violent + 1))

# create scaled DVs (to account for different pop. sizes and density in blocks)
mil_fin$larrest_sd <- mil_fin$larrests - (mean(mil_fin$larrests)/sd(mil_fin$larrests))
mil_fin$lmisdemeanors_sd <- mil_fin$lmisdemeanors - (mean(mil_fin$lmisdemeanors)/sd(mil_fin$lmisdemeanors))
mil_fin$lfelonies_sd <- mil_fin$lfelonies - (mean(mil_fin$lfelonies)/sd(mil_fin$lfelonies))
mil_fin$lnonviolent_sd <- mil_fin$lnonviolent - (mean(mil_fin$lnonviolent)/sd(mil_fin$lnonviolent))
mil_fin$lviolent_sd <- mil_fin$lviolent - (mean(mil_fin$lviolent)/sd(mil_fin$lviolent))
mil_fin$lsociety_sd <- mil_fin$lsociety - (mean(mil_fin$lsociety)/sd(mil_fin$lsociety))
mil_fin$lperson_sd <- mil_fin$lperson - (mean(mil_fin$lperson)/sd(mil_fin$lperson))
mil_fin$lproperty_sd <- mil_fin$lproperty - (mean(mil_fin$lproperty)/sd(mil_fin$lproperty))

# now create binned racial blv measures
mil_fin <- mil_fin %>% mutate(boundary_quart = quantile(mil_fin$p_race_white_blv, prob=.75, na.rm=TRUE),
                              boundary_quart_dummy = ifelse(p_race_white_blv > boundary_quart,1,0 ))


# For Seattle
## Load Seattle data
load("sea_final.RData")


# Set up DVs
# log and +1 to variables (pop already logged) 
sea_fin = sea_fin %>% 
  mutate(lmhhi = log(mhhi + 1),
         larrests = log(total_arrests + 1),
         lmisdemeanors = log(misdemeanor_arrests + 1),
         lfelonies = log(felony_arrests + 1),
         lnonviolent = log(nonviolent_arrests + 1),
         lviolent = log(violent_arrests + 1),
         lsociety = log(society_arrests + 1),
         lperson = log(person_arrests + 1),
         lproperty = log(property_arrests + 1),
         lcrime = log(crime_all + 1),
         lpropertycrime = log(crime_property + 1),
         lviolentcrime = log(crime_violent + 1))

# create scaled DVs (to account for different pop. sizes and density in blocks)
sea_fin$larrest_sd <- sea_fin$larrests - (mean(sea_fin$larrests)/sd(sea_fin$larrests))
sea_fin$lmisdemeanors_sd <- sea_fin$lmisdemeanors - (mean(sea_fin$lmisdemeanors)/sd(sea_fin$lmisdemeanors))
sea_fin$lfelonies_sd <- sea_fin$lfelonies - (mean(sea_fin$lfelonies)/sd(sea_fin$lfelonies))
sea_fin$lnonviolent_sd <- sea_fin$lnonviolent - (mean(sea_fin$lnonviolent)/sd(sea_fin$lnonviolent))
sea_fin$lviolent_sd <- sea_fin$lviolent - (mean(sea_fin$lviolent)/sd(sea_fin$lviolent))
sea_fin$lsociety_sd <- sea_fin$lsociety - (mean(sea_fin$lsociety)/sd(sea_fin$lsociety))
sea_fin$lperson_sd <- sea_fin$lperson - (mean(sea_fin$lperson)/sd(sea_fin$lperson))
sea_fin$lproperty_sd <- sea_fin$lproperty - (mean(sea_fin$lproperty)/sd(sea_fin$lproperty))

# now create binned racial blv measures
sea_fin <- sea_fin %>% mutate(boundary_quart = quantile(sea_fin$p_race_white_blv, prob=.75, na.rm=TRUE),
                              boundary_quart_dummy = ifelse(p_race_white_blv > boundary_quart,1,0 ))



# now run interaction models with binned racial boundary value
form_list = 
  list(
    'lmisdemeanors_sd ~ boundary_quart_dummy * p_race_white + ses_blv + lpop + hhi + lmhhi + age_15_35_male  + 
    pown + p_poverty + p_emp_unemployed + pcol + lpropertycrime + lviolentcrime + com_density + phys_bound',
    'lfelonies_sd ~ boundary_quart_dummy * p_race_white + ses_blv + lpop + hhi + lmhhi + age_15_35_male  + 
    pown + p_poverty + p_emp_unemployed + pcol + lpropertycrime + lviolentcrime + com_density + phys_bound'
  )

hetdf_list = list(form_list)
hetdf_lab = c("Misdemeanors","Felonies")

out_mod_list = as.list(rep(NA, length(form_list)))
out_df_list = as.list(rep(NA, length(form_list)))

for (i in 1:length(form_list)) {
  
  print(paste0("Iteration ", i))
  
  out_mod_list[[i]] = lm_robust(as.formula(form_list[[i]]), atl_fin)
  
  fakedata = 
    atl_fin[, names(out_mod_list[[i]]$coefficients)[2:(length(names(out_mod_list[[i]]$coefficients))-1)]] %>% 
    filter(lpop > 0) %>% 
    as.data.frame %>% 
    mutate(geometry = NULL) %>% 
    apply(X = ., MARGIN = 2, FUN = function(x) mean(x, na.rm = TRUE)) %>% 
    t %>% 
    as.data.frame %>% 
    slice(rep(1:n(), each = 22)) %>% 
    mutate(p_race_white = rep(seq(from = 0, to = 1, by = .1), 2),
           boundary_quart_dummy = c(rep(0, 11), rep(1, 11)))
  
  out_df_list[[i]] = data.frame(
    
    est = predict(object = out_mod_list[[i]], newdata = fakedata, se.fit = TRUE)$fit,
    se = predict(object = out_mod_list[[i]], newdata = fakedata, se.fit = TRUE)$se.fit,
    p_race_white = rep(seq(from = 0, to = 1, by = .1), 2),
    boundary_quart_dummy = c(rep(0, 11), rep(1, 11))
    
  ) %>% 
    mutate(dataset = hetdf_lab[i])
  
}

df_to_plot = out_df_list %>%
  do.call(rbind.data.frame, .) %>%
  mutate(boundary_quart_dummy = factor(boundary_quart_dummy, labels = c("No Boundary", "Boundary")))

df_to_plot = df_to_plot %>%
  mutate(city = "Atlanta")

# create interaction plots

atl_racial_blv_int = df_to_plot %>%
  ggplot() +
  geom_point(aes(x = p_race_white, y = est, col = boundary_quart_dummy),
             position = position_dodge(.075),
             size = 2) +
  geom_errorbar(aes(x = p_race_white,
                    ymin = est - 1.96 * se,
                    ymax = est + 1.96 * se,
                    col = boundary_quart_dummy),
                width = 0,
                size = .4,
                position = position_dodge(.075)) +
  geom_errorbar(aes(x = p_race_white,
                    ymin = est - 1.645 * se,
                    ymax = est + 1.645 * se,
                    col = boundary_quart_dummy),
                width = 0,
                size = .6,
                position = position_dodge(.075)) +
  scale_color_grey(start = .6, end = 0) +
  labs(x = "% White",
       y = "Predicted Value",
       col = "Race Boundary") +
  facet_grid2(city ~ dataset, scales = "free_x", independent = "x") + 
  theme_bw(base_size=12,base_family="Times")+
  theme(panel.border=element_rect(fill=NA, colour=NA), 
        legend.title = element_blank(),
        legend.position="none",
        panel.grid.minor=element_line(colour=NA)) +
  theme(strip.background = element_rect(fill="white")) + 
  theme(panel.grid.major = element_blank(), panel.grid.minor = element_blank(),
        axis.title.x=element_blank())


# For Austin
# load final data
load("aus_final.RData")


# Set up DVs
# log and +1 to variables (pop already logged)
aus_fin = aus_fin %>% 
  mutate(lmhhi = log(mhhi + 1),
         larrests = log(total_arrests + 1),
         lmisdemeanors = log(misdemeanor_arrests + 1),
         lfelonies = log(felony_arrests + 1),
         lnonviolent = log(nonviolent_arrests + 1),
         lviolent = log(violent_arrests + 1),
         lsociety = log(society_arrests + 1),
         lperson = log(person_arrests + 1),
         lproperty = log(property_arrests + 1),
         lcrime = log(crime_all + 1),
         lpropertycrime = log(crime_property + 1),
         lviolentcrime = log(crime_violent + 1))

# create scaled DVs (to account for different pop. sizes and density in blocks)
aus_fin$larrest_sd <- aus_fin$larrests - (mean(aus_fin$larrests)/sd(aus_fin$larrests))
aus_fin$lmisdemeanors_sd <- aus_fin$lmisdemeanors - (mean(aus_fin$lmisdemeanors)/sd(aus_fin$lmisdemeanors))
aus_fin$lfelonies_sd <- aus_fin$lfelonies - (mean(aus_fin$lfelonies)/sd(aus_fin$lfelonies))
aus_fin$lnonviolent_sd <- aus_fin$lnonviolent - (mean(aus_fin$lnonviolent)/sd(aus_fin$lnonviolent))
aus_fin$lviolent_sd <- aus_fin$lviolent - (mean(aus_fin$lviolent)/sd(aus_fin$lviolent))
aus_fin$lsociety_sd <- aus_fin$lsociety - (mean(aus_fin$lsociety)/sd(aus_fin$lsociety))
aus_fin$lperson_sd <- aus_fin$lperson - (mean(aus_fin$lperson)/sd(aus_fin$lperson))
aus_fin$lproperty_sd <- aus_fin$lproperty - (mean(aus_fin$lproperty)/sd(aus_fin$lproperty))

# now create binned racial blv measures
aus_fin <- aus_fin %>% mutate(boundary_quart = quantile(aus_fin$p_race_white_blv, prob=.75, na.rm=TRUE),
                              boundary_quart_dummy = ifelse(p_race_white_blv > boundary_quart,1,0 ))


# now run interaction models with binned racial boundary value
form_list = 
  list(
    'lmisdemeanors_sd ~ boundary_quart_dummy * p_race_white + ses_blv + lpop + hhi + lmhhi + age_15_35_male  + pown + 
    p_poverty + p_emp_unemployed + pcol + lpropertycrime + lviolentcrime + com_density + phys_bound',
    'lfelonies_sd ~ boundary_quart_dummy * p_race_white + ses_blv + lpop + hhi + lmhhi + age_15_35_male  + pown + 
    p_poverty + p_emp_unemployed + pcol + lpropertycrime + lviolentcrime + com_density + phys_bound')

hetdf_list = list(form_list)
hetdf_lab = c( "Misdemeanors","Felonies")

out_mod_list = as.list(rep(NA, length(form_list)))
out_df_list = as.list(rep(NA, length(form_list)))

for (i in 1:length(form_list)) {
  
  print(paste0("Iteration ", i))
  
  out_mod_list[[i]] = lm_robust(as.formula(form_list[[i]]), aus_fin)
  
  fakedata = 
    aus_fin[, names(out_mod_list[[i]]$coefficients)[2:(length(names(out_mod_list[[i]]$coefficients))-1)]] %>% 
    filter(lpop > 0) %>% 
    as.data.frame %>% 
    mutate(geometry = NULL) %>% 
    apply(X = ., MARGIN = 2, FUN = function(x) mean(x, na.rm = TRUE)) %>% 
    t %>% 
    as.data.frame %>% 
    slice(rep(1:n(), each = 22)) %>% 
    mutate(p_race_white = rep(seq(from = 0, to = 1, by = .1), 2),
           boundary_quart_dummy = c(rep(0, 11), rep(1, 11)))
  
  out_df_list[[i]] = data.frame(
    
    est = predict(object = out_mod_list[[i]], newdata = fakedata, se.fit = TRUE)$fit,
    se = predict(object = out_mod_list[[i]], newdata = fakedata, se.fit = TRUE)$se.fit,
    p_race_white = rep(seq(from = 0, to = 1, by = .1), 2),
    boundary_quart_dummy = c(rep(0, 11), rep(1, 11))
    
  ) %>% 
    mutate(dataset = hetdf_lab[i])
  
}

df_to_plot = out_df_list %>%
  do.call(rbind.data.frame, .) %>%
  mutate(boundary_quart_dummy = factor(boundary_quart_dummy, labels = c("No Boundary", "Boundary")))

df_to_plot = df_to_plot %>%
  mutate(city = "Austin")

# create interaction plots

aus_racial_blv_int = df_to_plot %>%
  ggplot() +
  geom_point(aes(x = p_race_white, y = est, col = boundary_quart_dummy),
             position = position_dodge(.075),
             size = 2) +
  geom_errorbar(aes(x = p_race_white,
                    ymin = est - 1.96 * se,
                    ymax = est + 1.96 * se,
                    col = boundary_quart_dummy),
                width = 0,
                size = .4,
                position = position_dodge(.075)) +
  geom_errorbar(aes(x = p_race_white,
                    ymin = est - 1.645 * se,
                    ymax = est + 1.645 * se,
                    col = boundary_quart_dummy),
                width = 0,
                size = .6,
                position = position_dodge(.075)) +
  scale_color_grey(start = .6, end = 0) +
  labs(x = "% White",
       y = "Predicted Value",
       col = "Race Boundary") +
  facet_grid2(city ~ dataset, scales = "free_x", independent = "x") + 
  theme_bw(base_size=12,base_family="Times") +
  theme(panel.border=element_rect(fill=NA, colour=NA), 
        legend.title = element_blank(),
        legend.position="none", # remove legend
        panel.grid.minor=element_line(colour=NA)) +
  theme(strip.background = element_rect(fill="white")) + # remove grey from background
  theme(panel.grid.major = element_blank(), panel.grid.minor = element_blank(), # remove gridlines from background
        axis.title.x=element_blank(), strip.text.x = element_blank()) # remove x axis title and x labels for plots


# For Boston
# load final data
load("bos_final.RData")


# Set up DVs
# log and +1 to variables (pop already logged)
bos_fin = bos_fin %>% 
  mutate(lmhhi = log(mhhi + 1),
         larrests = log(total_arrests + 1),
         lmisdemeanors = log(misdemeanor_arrests + 1),
         lfelonies = log(felony_arrests + 1),
         lnonviolent = log(nonviolent_arrests + 1),
         lviolent = log(violent_arrests + 1),
         lsociety = log(society_arrests + 1),
         lperson = log(person_arrests + 1),
         lproperty = log(property_arrests + 1),
         lcrime = log(crime + 1))



# create scaled DVs (to account for different pop. sizes and density in blocks)
bos_fin$larrests_sd <- bos_fin$larrests - (mean(bos_fin$larrests)/sd(bos_fin$larrests))
bos_fin$lmisdemeanors_sd <- bos_fin$lmisdemeanors - (mean(bos_fin$lmisdemeanors)/sd(bos_fin$lmisdemeanors))
bos_fin$lfelonies_sd <- bos_fin$lfelonies - (mean(bos_fin$lfelonies)/sd(bos_fin$lfelonies))
bos_fin$lnonviolent_sd <- bos_fin$lnonviolent - (mean(bos_fin$lnonviolent)/sd(bos_fin$lnonviolent))
bos_fin$lviolent_sd <- bos_fin$lviolent - (mean(bos_fin$lviolent)/sd(bos_fin$lviolent))
bos_fin$lsociety_sd <- bos_fin$lsociety - (mean(bos_fin$lsociety)/sd(bos_fin$lsociety))
bos_fin$lperson_sd <- bos_fin$lperson - (mean(bos_fin$lperson)/sd(bos_fin$lperson))
bos_fin$lproperty_sd <- bos_fin$lproperty - (mean(bos_fin$lproperty)/sd(bos_fin$lproperty))


# now create binned racial blv measures
bos_fin <- bos_fin %>% mutate(boundary_quart = quantile(bos_fin$p_race_white_blv, prob=.75, na.rm=TRUE),
                              boundary_quart_dummy = ifelse(p_race_white_blv > boundary_quart,1,0 ))


# now run interaction models with binned racial boundary value
form_list = 
  list(
    'lmisdemeanors_sd ~ boundary_quart_dummy * p_race_white + ses_blv + lpop + hhi + lmhhi + age_15_35_male  + pown + 
    p_poverty + p_emp_unemployed + pcol + lcrime + com_density + phys_bound',
    'lfelonies_sd ~ boundary_quart_dummy * p_race_white + ses_blv + lpop + hhi + lmhhi + age_15_35_male  + pown + 
    p_poverty + p_emp_unemployed + pcol  + lcrime + com_density + phys_bound')

hetdf_list = list(form_list)
hetdf_lab = c("Misdemeanors","Felonies")

out_mod_list = as.list(rep(NA, length(form_list)))
out_df_list = as.list(rep(NA, length(form_list)))

for (i in 1:length(form_list)) {
  
  print(paste0("Iteration ", i))
  
  out_mod_list[[i]] = lm_robust(as.formula(form_list[[i]]), bos_fin)
  
  fakedata = 
    bos_fin[, names(out_mod_list[[i]]$coefficients)[2:(length(names(out_mod_list[[i]]$coefficients))-1)]] %>% 
    filter(lpop > 0) %>% 
    as.data.frame %>% 
    mutate(geometry = NULL) %>% 
    apply(X = ., MARGIN = 2, FUN = function(x) mean(x, na.rm = TRUE)) %>% 
    t %>% 
    as.data.frame %>% 
    slice(rep(1:n(), each = 22)) %>% 
    mutate(p_race_white = rep(seq(from = 0, to = 1, by = .1), 2),
           boundary_quart_dummy = c(rep(0, 11), rep(1, 11)))
  
  out_df_list[[i]] = data.frame(
    
    est = predict(object = out_mod_list[[i]], newdata = fakedata, se.fit = TRUE)$fit,
    se = predict(object = out_mod_list[[i]], newdata = fakedata, se.fit = TRUE)$se.fit,
    p_race_white = rep(seq(from = 0, to = 1, by = .1), 2),
    boundary_quart_dummy = c(rep(0, 11), rep(1, 11))
    
  ) %>% 
    mutate(dataset = hetdf_lab[i])
  
}

df_to_plot = out_df_list %>%
  do.call(rbind.data.frame, .) %>%
  mutate(boundary_quart_dummy = factor(boundary_quart_dummy, labels = c("No Boundary", "Boundary")))

df_to_plot = df_to_plot %>%
  mutate(city = "Boston")

# create interaction plots

bos_racial_blv_int = df_to_plot %>%
  ggplot() +
  geom_point(aes(x = p_race_white, y = est, col = boundary_quart_dummy),
             position = position_dodge(.075),
             size = 2) +
  geom_errorbar(aes(x = p_race_white,
                    ymin = est - 1.96 * se,
                    ymax = est + 1.96 * se,
                    col = boundary_quart_dummy),
                width = 0,
                size = .4,
                position = position_dodge(.075)) +
  geom_errorbar(aes(x = p_race_white,
                    ymin = est - 1.645 * se,
                    ymax = est + 1.645 * se,
                    col = boundary_quart_dummy),
                width = 0,
                size = .6,
                position = position_dodge(.075)) +
  scale_color_grey(start = .6, end = 0) +
  labs(x = "% White",
       y = "Predicted Value",
       col = "Race Boundary") +
  facet_grid2(city ~ dataset, scales = "free_x", independent = "x")  +  
  theme_bw(base_size=12,base_family="Times") +
  theme(panel.border=element_rect(fill=NA, colour=NA), 
        legend.title = element_blank(),
        legend.position="none", # remove legend
        panel.grid.minor=element_line(colour=NA)) +
  theme(strip.background = element_rect(fill="white")) + # remove grey from background
  theme(panel.grid.major = element_blank(), panel.grid.minor = element_blank(), # remove gridlines from background
        axis.title.x=element_blank(), strip.text.x = element_blank()) # remove x axis title and x labels for plots


# For Chicago
# load final data
load("chi_final.RData")


# Set up DVs
# log and +1 to variables (pop already logged)
chi_fin = chi_fin %>% 
  mutate(lmhhi = log(mhhi + 1),
         larrests = log(total_arrests + 1),
         lmisdemeanors = log(misdemeanor_arrests + 1),
         lfelonies = log(felony_arrests + 1),
         lnonviolent = log(nonviolent_arrests + 1),
         lviolent = log(violent_arrests + 1),
         lsociety = log(society_arrests + 1),
         lperson = log(person_arrests + 1),
         lproperty = log(property_arrests + 1),
         lcrime = log(crime_all + 1),
         lpropertycrime = log(property_crime + 1),
         lviolentcrime = log(violent_crime + 1))


# create scaled DVs (to account for different pop. sizes and density in blocks)
chi_fin$larrest_sd <- chi_fin$larrests - (mean(chi_fin$larrests)/sd(chi_fin$larrests))
chi_fin$lmisdemeanors_sd <- chi_fin$lmisdemeanors - (mean(chi_fin$lmisdemeanors)/sd(chi_fin$lmisdemeanors))
chi_fin$lfelonies_sd <- chi_fin$lfelonies - (mean(chi_fin$lfelonies)/sd(chi_fin$lfelonies))
chi_fin$lnonviolent_sd <- chi_fin$lnonviolent - (mean(chi_fin$lnonviolent)/sd(chi_fin$lnonviolent))
chi_fin$lviolent_sd <- chi_fin$lviolent - (mean(chi_fin$lviolent)/sd(chi_fin$lviolent))
chi_fin$lsociety_sd <- chi_fin$lsociety - (mean(chi_fin$lsociety)/sd(chi_fin$lsociety))
chi_fin$lperson_sd <- chi_fin$lperson - (mean(chi_fin$lperson)/sd(chi_fin$lperson))
chi_fin$lproperty_sd <- chi_fin$lproperty - (mean(chi_fin$lproperty)/sd(chi_fin$lproperty))

# now create binned racial blv measures
chi_fin <- chi_fin %>% mutate(boundary_quart = quantile(chi_fin$p_race_white_blv, prob=.75, na.rm=TRUE),
                              boundary_quart_dummy = ifelse(p_race_white_blv > boundary_quart,1,0 ))


# now run interaction models with binned racial boundary value
form_list = 
  list(
    'lmisdemeanors_sd ~ boundary_quart_dummy * p_race_white + ses_blv + lpop + hhi + lmhhi + age_15_35_male  + pown + 
    p_poverty + p_emp_unemployed + pcol + lpropertycrime + lviolentcrime + com_density + phys_bound',
    'lfelonies_sd ~ boundary_quart_dummy * p_race_white + ses_blv + lpop + hhi + lmhhi + age_15_35_male  + pown + 
    p_poverty + p_emp_unemployed + pcol + lpropertycrime + lviolentcrime + com_density + phys_bound')

hetdf_list = list(form_list)
hetdf_lab = c("Misdemeanors","Felonies")

out_mod_list = as.list(rep(NA, length(form_list)))
out_df_list = as.list(rep(NA, length(form_list)))

for (i in 1:length(form_list)) {
  
  print(paste0("Iteration ", i))
  
  out_mod_list[[i]] = lm_robust(as.formula(form_list[[i]]), chi_fin)
  
  fakedata = 
    chi_fin[, names(out_mod_list[[i]]$coefficients)[2:(length(names(out_mod_list[[i]]$coefficients))-1)]] %>% 
    filter(lpop > 0) %>% 
    as.data.frame %>% 
    mutate(geometry = NULL) %>% 
    apply(X = ., MARGIN = 2, FUN = function(x) mean(x, na.rm = TRUE)) %>% 
    t %>% 
    as.data.frame %>% 
    slice(rep(1:n(), each = 22)) %>% 
    mutate(p_race_white = rep(seq(from = 0, to = 1, by = .1), 2),
           boundary_quart_dummy = c(rep(0, 11), rep(1, 11)))
  
  out_df_list[[i]] = data.frame(
    
    est = predict(object = out_mod_list[[i]], newdata = fakedata, se.fit = TRUE)$fit,
    se = predict(object = out_mod_list[[i]], newdata = fakedata, se.fit = TRUE)$se.fit,
    p_race_white = rep(seq(from = 0, to = 1, by = .1), 2),
    boundary_quart_dummy = c(rep(0, 11), rep(1, 11))
    
  ) %>% 
    mutate(dataset = hetdf_lab[i])
  
}

df_to_plot = out_df_list %>%
  do.call(rbind.data.frame, .) %>%
  mutate(boundary_quart_dummy = factor(boundary_quart_dummy, labels = c("No Boundary", "Boundary")))

df_to_plot = df_to_plot %>%
  mutate(city = "Chicago")

# create interaction plots

chi_racial_blv_int = df_to_plot %>%
  ggplot() +
  geom_point(aes(x = p_race_white, y = est, col = boundary_quart_dummy),
             position = position_dodge(.075),
             size = 2) +
  geom_errorbar(aes(x = p_race_white,
                    ymin = est - 1.96 * se,
                    ymax = est + 1.96 * se,
                    col = boundary_quart_dummy),
                width = 0,
                size = .4,
                position = position_dodge(.075)) +
  geom_errorbar(aes(x = p_race_white,
                    ymin = est - 1.645 * se,
                    ymax = est + 1.645 * se,
                    col = boundary_quart_dummy),
                width = 0,
                size = .6,
                position = position_dodge(.075)) +
  scale_color_grey(start = .6, end = 0) +
  labs(x = "% White",
       y = "Predicted Value",
       col = "Race Boundary") +
  facet_grid2(city ~ dataset, scales = "free_x", independent = "x") + 
  theme_bw(base_size=12,base_family="Times") +
  theme(panel.border=element_rect(fill=NA, colour=NA), 
        legend.title = element_blank(),
        legend.position="none", # remove legend
        panel.grid.minor=element_line(colour=NA)) +
  theme(strip.background = element_rect(fill="white")) + # remove grey from background
  theme(panel.grid.major = element_blank(), panel.grid.minor = element_blank(), # remove gridlines from background
        axis.title.x=element_blank(), strip.text.x = element_blank()) # remove x axis title and x labels for plots


# For Louisville
# load final data
load("lou_final.RData")

# Set up DVs
# log and +1 to variables (pop already logged) (no misdemeanors vs felonies for louisville)
lou_fin = lou_fin %>% 
  mutate(lmhhi = log(mhhi + 1),
         larrests = log(total_arrests + 1),
         lnonviolent = log(nonviolent_arrests + 1),
         lsociety = log(society_arrests + 1),
         lcrime = log(crime_all + 1),
         lpropertycrime = log(crime_property+1),
         lviolentcrime = log(crime_violent+1),
         lviolent = log(violent_arrests + 1),
         lperson = log(person_arrests + 1),
         lproperty = log(property_arrests + 1))

# create scaled DVs (to account for different pop. sizes and density in blocks)
lou_fin$larrest_sd <- lou_fin$larrests - (mean(lou_fin$larrests)/sd(lou_fin$larrests))
# lou_fin$lmisdemeanors_sd <- lou_fin$lmisdemeanors - (mean(lou_fin$lmisdemeanors)/sd(lou_fin$lmisdemeanors))
lou_fin$lnonviolent_sd <- lou_fin$lnonviolent - (mean(lou_fin$lnonviolent)/sd(lou_fin$lnonviolent))
lou_fin$lsociety_sd <- lou_fin$lsociety - (mean(lou_fin$lsociety)/sd(lou_fin$lsociety))
# lou_fin$lfelonies_sd <- lou_fin$lfelonies - (mean(lou_fin$lfelonies)/sd(lou_fin$lfelonies))
lou_fin$lviolent_sd <- lou_fin$lviolent - (mean(lou_fin$lviolent)/sd(lou_fin$lviolent))
lou_fin$lperson_sd <- lou_fin$lperson - (mean(lou_fin$lperson)/sd(lou_fin$lperson))
lou_fin$lproperty_sd <- lou_fin$lproperty - (mean(lou_fin$lproperty)/sd(lou_fin$lproperty))
# now create binned racial blv measures
lou_fin <- lou_fin %>% mutate(boundary_quart = quantile(lou_fin$p_race_white_blv, prob=.75, na.rm=TRUE),
                              boundary_quart_dummy = ifelse(p_race_white_blv > boundary_quart,1,0 ))


# now run interaction models with binned racial boundary value
form_list = 
  list(
    'lsociety_sd ~ boundary_quart_dummy * p_race_white + ses_blv + lpop + hhi + lmhhi + age_15_35_male  + pown + 
    p_poverty + p_emp_unemployed + pcol + lpropertycrime + lviolentcrime + com_density + phys_bound',
    'lperson_sd ~ boundary_quart_dummy * p_race_white + ses_blv + lpop + hhi + lmhhi + age_15_35_male  + pown + 
    p_poverty + p_emp_unemployed + pcol + lpropertycrime + lviolentcrime + com_density + phys_bound',
    'lproperty_sd ~ boundary_quart_dummy * p_race_white + ses_blv + lpop + hhi + lmhhi + age_15_35_male  + pown + 
    p_poverty + p_emp_unemployed + pcol + lpropertycrime + lviolentcrime + com_density + phys_bound'
  )

hetdf_list = list(form_list)
hetdf_lab = c("Against Society",
              "Against Persons","Against Property")

out_mod_list = as.list(rep(NA, length(form_list)))
out_df_list = as.list(rep(NA, length(form_list)))

for (i in 1:length(form_list)) {
  
  print(paste0("Iteration ", i))
  
  out_mod_list[[i]] = lm_robust(as.formula(form_list[[i]]), lou_fin)
  
  fakedata = 
    lou_fin[, names(out_mod_list[[i]]$coefficients)[2:(length(names(out_mod_list[[i]]$coefficients))-1)]] %>% 
    filter(lpop > 0) %>% 
    as.data.frame %>% 
    mutate(geometry = NULL) %>% 
    apply(X = ., MARGIN = 2, FUN = function(x) mean(x, na.rm = TRUE)) %>% 
    t %>% 
    as.data.frame %>% 
    slice(rep(1:n(), each = 22)) %>% 
    mutate(p_race_white = rep(seq(from = 0, to = 1, by = .1), 2),
           boundary_quart_dummy = c(rep(0, 11), rep(1, 11)))
  
  out_df_list[[i]] = data.frame(
    
    est = predict(object = out_mod_list[[i]], newdata = fakedata, se.fit = TRUE)$fit,
    se = predict(object = out_mod_list[[i]], newdata = fakedata, se.fit = TRUE)$se.fit,
    p_race_white = rep(seq(from = 0, to = 1, by = .1), 2),
    boundary_quart_dummy = c(rep(0, 11), rep(1, 11))
    
  ) %>% 
    mutate(dataset = hetdf_lab[i])
  
}

df_to_plot = out_df_list %>%
  do.call(rbind.data.frame, .) %>%
  mutate(boundary_quart_dummy = factor(boundary_quart_dummy, labels = c("No Boundary", "Boundary")))

# create interaction plots

lou_racial_blv_int = df_to_plot %>%
  ggplot() +
  geom_point(aes(x = p_race_white, y = est, col = boundary_quart_dummy),
             position = position_dodge(.075),
             size = 2) +
  geom_errorbar(aes(x = p_race_white,
                    ymin = est - 1.96 * se,
                    ymax = est + 1.96 * se,
                    col = boundary_quart_dummy),
                width = 0,
                size = .4,
                position = position_dodge(.075)) +
  geom_errorbar(aes(x = p_race_white,
                    ymin = est - 1.645 * se,
                    ymax = est + 1.645 * se,
                    col = boundary_quart_dummy),
                width = 0,
                size = .6,
                position = position_dodge(.075)) +
  scale_color_grey(start = .6, end = 0) +
  labs(x = "% White",
       y = "Predicted Value",
       col = "Race Boundary",
       title = "Louisville Moderation Analysis") +
  facet_wrap(~ dataset, scales = "free") + theme_bw(base_size=12,base_family="Times")+
  theme(panel.border=element_rect(fill=NA, colour=NA), 
        legend.title = element_blank(),
        legend.position="bottom",
        panel.grid.minor=element_line(colour=NA)) +
  theme(strip.background = element_rect(fill="white")) + 
  theme(panel.grid.major = element_blank(), panel.grid.minor = element_blank())


ggsave(width = 8, height = 6, plot = lou_racial_blv_int, filename = "lou_racial_blv_int.png")


# For Milwaukee
# load final data
load("mil_final.RData")


# Set up DVs
# log and +1 to variables (pop already logged) 
mil_fin = mil_fin %>% 
  mutate(lmhhi = log(mhhi + 1),
         larrests = log(total_arrests + 1),
         lmisdemeanors = log(misdemeanor_arrests + 1),
         lfelonies = log(felony_arrests + 1),
         lnonviolent = log(nonviolent_arrests + 1),
         lviolent = log(violent_arrests + 1),
         lsociety = log(society_arrests + 1),
         lperson = log(person_arrests + 1),
         lproperty = log(property_arrests + 1),
         lcrime = log(crime_all + 1),
         lpropertycrime = log(crime_property + 1),
         lviolentcrime = log(crime_violent + 1))

# create scaled DVs (to account for different pop. sizes and density in blocks)
mil_fin$larrest_sd <- mil_fin$larrests - (mean(mil_fin$larrests)/sd(mil_fin$larrests))
mil_fin$lmisdemeanors_sd <- mil_fin$lmisdemeanors - (mean(mil_fin$lmisdemeanors)/sd(mil_fin$lmisdemeanors))
mil_fin$lfelonies_sd <- mil_fin$lfelonies - (mean(mil_fin$lfelonies)/sd(mil_fin$lfelonies))
mil_fin$lnonviolent_sd <- mil_fin$lnonviolent - (mean(mil_fin$lnonviolent)/sd(mil_fin$lnonviolent))
mil_fin$lviolent_sd <- mil_fin$lviolent - (mean(mil_fin$lviolent)/sd(mil_fin$lviolent))
mil_fin$lsociety_sd <- mil_fin$lsociety - (mean(mil_fin$lsociety)/sd(mil_fin$lsociety))
mil_fin$lperson_sd <- mil_fin$lperson - (mean(mil_fin$lperson)/sd(mil_fin$lperson))
mil_fin$lproperty_sd <- mil_fin$lproperty - (mean(mil_fin$lproperty)/sd(mil_fin$lproperty))

# now create binned racial blv measures
mil_fin <- mil_fin %>% mutate(boundary_quart = quantile(mil_fin$p_race_white_blv, prob=.75, na.rm=TRUE),
                              boundary_quart_dummy = ifelse(p_race_white_blv > boundary_quart,1,0 ))

# now run interaction models with binned racial boundary value
form_list = 
  list(
    'lmisdemeanors_sd ~ boundary_quart_dummy * p_race_white + ses_blv + lpop + hhi + lmhhi + age_15_35_male  + pown + 
    p_poverty + p_emp_unemployed + pcol + lpropertycrime + lviolentcrime + com_density + phys_bound',
    'lfelonies_sd ~ boundary_quart_dummy * p_race_white + ses_blv + lpop + hhi + lmhhi + age_15_35_male  + pown + 
    p_poverty + p_emp_unemployed + pcol + lpropertycrime + lviolentcrime + com_density + phys_bound')

hetdf_list = list(form_list)
hetdf_lab = c("Misdemeanors","Felonies")

out_mod_list = as.list(rep(NA, length(form_list)))
out_df_list = as.list(rep(NA, length(form_list)))

for (i in 1:length(form_list)) {
  
  print(paste0("Iteration ", i))
  
  out_mod_list[[i]] = lm_robust(as.formula(form_list[[i]]), mil_fin)
  
  fakedata = 
    mil_fin[, names(out_mod_list[[i]]$coefficients)[2:(length(names(out_mod_list[[i]]$coefficients))-1)]] %>% 
    filter(lpop > 0) %>% 
    as.data.frame %>% 
    mutate(geometry = NULL) %>% 
    apply(X = ., MARGIN = 2, FUN = function(x) mean(x, na.rm = TRUE)) %>% 
    t %>% 
    as.data.frame %>% 
    slice(rep(1:n(), each = 22)) %>% 
    mutate(p_race_white = rep(seq(from = 0, to = 1, by = .1), 2),
           boundary_quart_dummy = c(rep(0, 11), rep(1, 11)))
  
  out_df_list[[i]] = data.frame(
    
    est = predict(object = out_mod_list[[i]], newdata = fakedata, se.fit = TRUE)$fit,
    se = predict(object = out_mod_list[[i]], newdata = fakedata, se.fit = TRUE)$se.fit,
    p_race_white = rep(seq(from = 0, to = 1, by = .1), 2),
    boundary_quart_dummy = c(rep(0, 11), rep(1, 11))
    
  ) %>% 
    mutate(dataset = hetdf_lab[i])
  
}

df_to_plot = out_df_list %>%
  do.call(rbind.data.frame, .) %>%
  mutate(boundary_quart_dummy = factor(boundary_quart_dummy, labels = c("No Boundary", "Boundary")))

df_to_plot = df_to_plot %>%
  mutate(city = "Milwaukee")

# create interaction plots

mil_racial_blv_int = df_to_plot %>%
  ggplot() +
  geom_point(aes(x = p_race_white, y = est, col = boundary_quart_dummy),
             position = position_dodge(.075),
             size = 2) +
  geom_errorbar(aes(x = p_race_white,
                    ymin = est - 1.96 * se,
                    ymax = est + 1.96 * se,
                    col = boundary_quart_dummy),
                width = 0,
                size = .4,
                position = position_dodge(.075)) +
  geom_errorbar(aes(x = p_race_white,
                    ymin = est - 1.645 * se,
                    ymax = est + 1.645 * se,
                    col = boundary_quart_dummy),
                width = 0,
                size = .6,
                position = position_dodge(.075)) +
  scale_color_grey(start = .6, end = 0) +
  labs(x = "% White",
       y = "Predicted Value",
       col = "Race Boundary") +
  facet_grid2(city ~ dataset, scales = "free_x", independent = "x") + 
  theme_bw(base_size=12,base_family="Times") +
  theme(panel.border=element_rect(fill=NA, colour=NA), 
        legend.title = element_blank(),
        legend.position="none", # remove legend
        panel.grid.minor=element_line(colour=NA)) +
  theme(strip.background = element_rect(fill="white")) + # remove grey from background
  theme(panel.grid.major = element_blank(), panel.grid.minor = element_blank(), # remove gridlines from background
        axis.title.x=element_blank(), strip.text.x = element_blank()) # remove x axis title and x labels for plots

# For Seattle
# load final data
load("sea_final.RData")


# Set up DVs
# log and +1 to variables (pop already logged) 
sea_fin = sea_fin %>% 
  mutate(lmhhi = log(mhhi + 1),
         larrests = log(total_arrests + 1),
         lmisdemeanors = log(misdemeanor_arrests + 1),
         lfelonies = log(felony_arrests + 1),
         lnonviolent = log(nonviolent_arrests + 1),
         lviolent = log(violent_arrests + 1),
         lsociety = log(society_arrests + 1),
         lperson = log(person_arrests + 1),
         lproperty = log(property_arrests + 1),
         lcrime = log(crime_all + 1),
         lpropertycrime = log(crime_property + 1),
         lviolentcrime = log(crime_violent + 1))

# create scaled DVs (to account for different pop. sizes and density in blocks)
sea_fin$larrest_sd <- sea_fin$larrests - (mean(sea_fin$larrests)/sd(sea_fin$larrests))
sea_fin$lmisdemeanors_sd <- sea_fin$lmisdemeanors - (mean(sea_fin$lmisdemeanors)/sd(sea_fin$lmisdemeanors))
sea_fin$lfelonies_sd <- sea_fin$lfelonies - (mean(sea_fin$lfelonies)/sd(sea_fin$lfelonies))
sea_fin$lnonviolent_sd <- sea_fin$lnonviolent - (mean(sea_fin$lnonviolent)/sd(sea_fin$lnonviolent))
sea_fin$lviolent_sd <- sea_fin$lviolent - (mean(sea_fin$lviolent)/sd(sea_fin$lviolent))
sea_fin$lsociety_sd <- sea_fin$lsociety - (mean(sea_fin$lsociety)/sd(sea_fin$lsociety))
sea_fin$lperson_sd <- sea_fin$lperson - (mean(sea_fin$lperson)/sd(sea_fin$lperson))
sea_fin$lproperty_sd <- sea_fin$lproperty - (mean(sea_fin$lproperty)/sd(sea_fin$lproperty))

# now create binned racial blv measures
sea_fin <- sea_fin %>% mutate(boundary_quart = quantile(sea_fin$p_race_white_blv, prob=.75, na.rm=TRUE),
                              boundary_quart_dummy = ifelse(p_race_white_blv > boundary_quart,1,0 ))

# now run interaction models with binned racial boundary value
form_list = 
  list(
    'lmisdemeanors_sd ~ boundary_quart_dummy * p_race_white + ses_blv + lpop + hhi + lmhhi + age_15_35_male  + pown + 
    p_poverty + p_emp_unemployed + pcol + lpropertycrime + lviolentcrime + com_density + phys_bound',
    'lfelonies_sd ~ boundary_quart_dummy * p_race_white + ses_blv + lpop + hhi + lmhhi + age_15_35_male  + pown + 
    p_poverty + p_emp_unemployed + pcol + lpropertycrime + lviolentcrime + com_density + phys_bound')

hetdf_list = list(form_list)
hetdf_lab = c("Misdemeanors","Felonies")

out_mod_list = as.list(rep(NA, length(form_list)))
out_df_list = as.list(rep(NA, length(form_list)))

for (i in 1:length(form_list)) {
  
  print(paste0("Iteration ", i))
  
  out_mod_list[[i]] = lm_robust(as.formula(form_list[[i]]), sea_fin)
  
  fakedata = 
    sea_fin[, names(out_mod_list[[i]]$coefficients)[2:(length(names(out_mod_list[[i]]$coefficients))-1)]] %>% 
    filter(lpop > 0) %>% 
    as.data.frame %>% 
    mutate(geometry = NULL) %>% 
    apply(X = ., MARGIN = 2, FUN = function(x) mean(x, na.rm = TRUE)) %>% 
    t %>% 
    as.data.frame %>% 
    slice(rep(1:n(), each = 22)) %>% 
    mutate(p_race_white = rep(seq(from = 0, to = 1, by = .1), 2),
           boundary_quart_dummy = c(rep(0, 11), rep(1, 11)))
  
  out_df_list[[i]] = data.frame(
    
    est = predict(object = out_mod_list[[i]], newdata = fakedata, se.fit = TRUE)$fit,
    se = predict(object = out_mod_list[[i]], newdata = fakedata, se.fit = TRUE)$se.fit,
    p_race_white = rep(seq(from = 0, to = 1, by = .1), 2),
    boundary_quart_dummy = c(rep(0, 11), rep(1, 11))
    
  ) %>% 
    mutate(dataset = hetdf_lab[i])
  
}

df_to_plot = out_df_list %>%
  do.call(rbind.data.frame, .) %>%
  mutate(boundary_quart_dummy = factor(boundary_quart_dummy, labels = c("No Boundary", "Boundary")))

df_to_plot = df_to_plot %>%
  mutate(city = "Seattle")

# create interaction plots

sea_racial_blv_int = df_to_plot %>%
  ggplot() +
  geom_point(aes(x = p_race_white, y = est, col = boundary_quart_dummy),
             position = position_dodge(.075),
             size = 2) +
  geom_errorbar(aes(x = p_race_white,
                    ymin = est - 1.96 * se,
                    ymax = est + 1.96 * se,
                    col = boundary_quart_dummy),
                width = 0,
                size = .4,
                position = position_dodge(.075)) +
  geom_errorbar(aes(x = p_race_white,
                    ymin = est - 1.645 * se,
                    ymax = est + 1.645 * se,
                    col = boundary_quart_dummy),
                width = 0,
                size = .6,
                position = position_dodge(.075)) +
  scale_color_grey(start = .6, end = 0) +
  labs(x = "% White",
       y = "Predicted Value",
       col = "Race Boundary") +
  facet_grid2(city ~ dataset, scales = "free_x", independent = "x")  + 
  theme_bw(base_size=12,base_family="Times") +
  theme(panel.border=element_rect(fill=NA, colour=NA), 
        legend.title = element_blank(),
        legend.position="bottom", # remove legend
        panel.grid.minor=element_line(colour=NA)) +
  theme(strip.background = element_rect(fill="white")) + # remove grey from background
  theme(panel.grid.major = element_blank(), panel.grid.minor = element_blank(), # remove gridlines from background
        strip.text.x = element_blank()) # remove x axis title and x labels for plots


# arrange all plots on one page (expect Louisville)

library(gridExtra)
library(cowplot)
all_cities_mod_plots <- plot_grid(atl_racial_blv_int, aus_racial_blv_int, bos_racial_blv_int,
                                  chi_racial_blv_int,mil_racial_blv_int,sea_racial_blv_int, 
                                  align = "v", nrow = 6, rel_heights = c(1/5, 1/6, 1/6, 1/6, 1/6, 1/5))

ggsave(plot = all_cities_mod_plots, filename = "figure5_plots.png", 
       width = 8, height = 14, bg = 'white')

